openclaw - 💡(How to fix) Fix [Bug]: /new causes model to monologue unprompted with leaked context from previous sessions (since 2026.4.24) [1 comments, 2 participants]

Bug type

Regression (worked before, now fails)

Beta release blocker

No

Summary

Since v2026.4.24 (continuing through 4.25 and 4.26), using the /new slash command causes the model to start monologuing unprompted — it generates responses without any user input, pulling random content fragments from previous conversations and producing nonsensical, incoherent rambling. This happens with any model (DeepSeek, MiniMax, etc.), not specific to one provider.

Steps to reproduce

1. Have an active conversation with the agent (any platform — WeChat/Weixin, Telegram, etc.)
2. Type /new to start a fresh session
3. Observe: the agent begins generating messages on its own, without waiting for user input
4. The content of these messages is clearly derived from previous conversations — random snippets, old topics, disconnected fragments — but delivered as incoherent monologue

Expected behavior

/new should create a clean, blank session. The agent should wait for the user's first message. No content from previous sessions should leak through.

Actual behavior

• Agent immediately starts generating multiple messages unprompted
• Content is garbled fragments from past conversations
• Resembles the model "continuing" an old conversation it thinks is still active
• Affects ALL models — provider/model-agnostic
• Only fix is to restart the gateway or kill the runaway session

OpenClaw version

2026.4.24 through 2026.4.26 (be8c246)

Operating system

Linux (Ubuntu)

Install method

npm global

Model

All models affected (DeepSeek V4, MiniMax M2.5/M2.7, etc.)

Provider / routing chain

WeChat/Weixin to OpenClaw gateway to model provider

Additional provider/model setup details

Multi-provider setup with credential pools (MiniMax CN, DeepSeek)

Logs, screenshots, and evidence

The model generates unprompted output like:

• Randomly re-discussing topics from conversations hours/days ago
• Fragments of old system prompts or memory notes
• Incoherent strings mixing multiple past conversation threads
• No user message was sent, the agent just starts talking to itself

Impact and severity

• Affected: All users who use /new to reset sessions
• Severity: High — makes /new unusable, requires gateway restart to recover
• Frequency: Consistent (happens on almost every /new since 4.24)
• Consequence: Core session management is broken; users cannot reliably start fresh sessions

Suspected cause

Looks like /new is not properly clearing session context — residual conversation fragments or memory artifacts from the old session are being injected into the new session prompt. Without a proper user message as anchor, the model "continues" the garbled context it receives, resulting in unprompted monologue.

Related to existing issues:

• #64168 (self-improvement note leaking on /new)
• #32296 (session context confusion)
• #64262 (garbled output)

These may share the same root cause of incomplete session context cleanup on /new.

Additional information

This is a regression — /new worked correctly before v4.24. The behavior started at 4.24 and has persisted through every subsequent version (4.25, 4.26).